Tag
介绍 Simular 的 Agent S4 论文：通过元智能体从任务轨迹中自动构建并进化计算机操作接口，使开源 Qwen3.8-27B 模型在 OSWorld V2 上达到 21.3%，超越 GPT-5.5 和 Claude Opus 4.8，且成本仅为后者的极小部分。
The paper introduces Agent S4, an autonomous system that evolves computer-use interfaces from task experience, achieving state-of-the-art open-source performance on OSWorld V2 (21.3%) and outperforming GPT-5.5 and Claude Opus 4.8 at a fraction of the cost.
CoEvo-RSI introduces a recursive self-improvement framework using bidirectional historical inheritance to align model capabilities with evolving agent harnesses, reducing brain-body mismatch. Using only SFT on rejection-sampled trajectories, it makes Qwen3.5-9B competitive with larger models, surpassing DeepSeek-V4-Pro and GLM-5.3 on agent benchmarks.
RSIFlow 提出一种递归自我改进框架，通过诊断任务执行反馈动态选择组件级改进工作流，并演化元智能体以优化编排能力，在六个基准上达到 SOTA 平均性能。
EAU 是一个元智能体框架，在训练多轮 LLM 智能体时协同演化用户模拟器、环境与强化学习配置，使训练系统随策略一同进化。在三个交互式智能体基准上，EAU 比最强基线整体提升 14.7–15.0%，在 ColBench 上任务完成率最高提升 36.7%。
JIT-Agent is an open-source 27B model that dynamically generates task-specific harnesses for AI agents, outperforming hand-built systems with improved token efficiency.
A new paper proposes using a meta-agent to monitor and intervene when an agent goes off track, potentially doubling autonomous task completion.
New paper proposes a meta-agent that monitors AI agents and corrects their mistakes, achieving ~2x more task completion autonomously.
This paper presents a novel blueprint for self-improving agents that combines scaffold editing and weight training through a meta-agent and feedback-agent, achieving a 14x speedup on a CUDA kernel for AlphaFold.
This paper introduces the Meta-Agent Challenge (MAC), a benchmark for evaluating AI models' ability to autonomously develop agent systems through iterative programming. Results show that current models rarely match human baselines and exhibit issues like reward hacking, highlighting gaps in self-improvement capabilities.